2 * mm/readahead.c - address_space-level file readahead.
4 * Copyright (C) 2002, Linus Torvalds
6 * 09Apr2002 Andrew Morton
10 #include <linux/kernel.h>
13 #include <linux/module.h>
14 #include <linux/blkdev.h>
15 #include <linux/backing-dev.h>
16 #include <linux/task_io_accounting_ops.h>
17 #include <linux/pagevec.h>
18 #include <linux/pagemap.h>
21 * Initialise a struct file's readahead state. Assumes that the caller has
25 file_ra_state_init(struct file_ra_state *ra, struct address_space *mapping)
27 ra->ra_pages = mapping->backing_dev_info->ra_pages;
30 EXPORT_SYMBOL_GPL(file_ra_state_init);
32 #define list_to_page(head) (list_entry((head)->prev, struct page, lru))
35 * see if a page needs releasing upon read_cache_pages() failure
36 * - the caller of read_cache_pages() may have set PG_private before calling,
37 * such as the NFS fs marking pages that are cached locally on disk, thus we
38 * need to give the fs a chance to clean up in the event of an error
40 static void read_cache_pages_invalidate_page(struct address_space *mapping,
43 if (PagePrivate(page)) {
44 if (!trylock_page(page))
46 page->mapping = mapping;
47 do_invalidatepage(page, 0);
51 page_cache_release(page);
55 * release a list of pages, invalidating them first if need be
57 static void read_cache_pages_invalidate_pages(struct address_space *mapping,
58 struct list_head *pages)
62 while (!list_empty(pages)) {
63 victim = list_to_page(pages);
64 list_del(&victim->lru);
65 read_cache_pages_invalidate_page(mapping, victim);
70 * read_cache_pages - populate an address space with some pages & start reads against them
71 * @mapping: the address_space
72 * @pages: The address of a list_head which contains the target pages. These
73 * pages have their ->index populated and are otherwise uninitialised.
74 * @filler: callback routine for filling a single page.
75 * @data: private data for the callback routine.
77 * Hides the details of the LRU cache etc from the filesystems.
79 int read_cache_pages(struct address_space *mapping, struct list_head *pages,
80 int (*filler)(void *, struct page *), void *data)
85 while (!list_empty(pages)) {
86 page = list_to_page(pages);
88 if (add_to_page_cache_lru(page, mapping,
89 page->index, GFP_KERNEL)) {
90 read_cache_pages_invalidate_page(mapping, page);
93 page_cache_release(page);
95 ret = filler(data, page);
97 read_cache_pages_invalidate_pages(mapping, pages);
100 task_io_account_read(PAGE_CACHE_SIZE);
105 EXPORT_SYMBOL(read_cache_pages);
107 static int read_pages(struct address_space *mapping, struct file *filp,
108 struct list_head *pages, unsigned nr_pages)
113 if (mapping->a_ops->readpages) {
114 ret = mapping->a_ops->readpages(filp, mapping, pages, nr_pages);
115 /* Clean up the remaining pages */
116 put_pages_list(pages);
120 for (page_idx = 0; page_idx < nr_pages; page_idx++) {
121 struct page *page = list_to_page(pages);
122 list_del(&page->lru);
123 if (!add_to_page_cache_lru(page, mapping,
124 page->index, GFP_KERNEL)) {
125 mapping->a_ops->readpage(filp, page);
127 page_cache_release(page);
135 * do_page_cache_readahead actually reads a chunk of disk. It allocates all
136 * the pages first, then submits them all for I/O. This avoids the very bad
137 * behaviour which would occur if page allocations are causing VM writeback.
138 * We really don't want to intermingle reads and writes like that.
140 * Returns the number of pages requested, or the maximum amount of I/O allowed.
142 * do_page_cache_readahead() returns -1 if it encountered request queue
146 __do_page_cache_readahead(struct address_space *mapping, struct file *filp,
147 pgoff_t offset, unsigned long nr_to_read,
148 unsigned long lookahead_size)
150 struct inode *inode = mapping->host;
152 unsigned long end_index; /* The last page we want to read */
153 LIST_HEAD(page_pool);
156 loff_t isize = i_size_read(inode);
161 end_index = ((isize - 1) >> PAGE_CACHE_SHIFT);
164 * Preallocate as many pages as we will need.
166 for (page_idx = 0; page_idx < nr_to_read; page_idx++) {
167 pgoff_t page_offset = offset + page_idx;
169 if (page_offset > end_index)
173 page = radix_tree_lookup(&mapping->page_tree, page_offset);
178 page = page_cache_alloc_cold(mapping);
181 page->index = page_offset;
182 list_add(&page->lru, &page_pool);
183 if (page_idx == nr_to_read - lookahead_size)
184 SetPageReadahead(page);
189 * Now start the IO. We ignore I/O errors - if the page is not
190 * uptodate then the caller will launch readpage again, and
191 * will then handle the error.
194 read_pages(mapping, filp, &page_pool, ret);
195 BUG_ON(!list_empty(&page_pool));
201 * Chunk the readahead into 2 megabyte units, so that we don't pin too much
204 int force_page_cache_readahead(struct address_space *mapping, struct file *filp,
205 pgoff_t offset, unsigned long nr_to_read)
209 if (unlikely(!mapping->a_ops->readpage && !mapping->a_ops->readpages))
215 unsigned long this_chunk = (2 * 1024 * 1024) / PAGE_CACHE_SIZE;
217 if (this_chunk > nr_to_read)
218 this_chunk = nr_to_read;
219 err = __do_page_cache_readahead(mapping, filp,
220 offset, this_chunk, 0);
226 offset += this_chunk;
227 nr_to_read -= this_chunk;
233 * This version skips the IO if the queue is read-congested, and will tell the
234 * block layer to abandon the readahead if request allocation would block.
236 * force_page_cache_readahead() will ignore queue congestion and will block on
239 int do_page_cache_readahead(struct address_space *mapping, struct file *filp,
240 pgoff_t offset, unsigned long nr_to_read)
242 if (bdi_read_congested(mapping->backing_dev_info))
245 return __do_page_cache_readahead(mapping, filp, offset, nr_to_read, 0);
249 * Given a desired number of PAGE_CACHE_SIZE readahead pages, return a
250 * sensible upper limit.
252 unsigned long max_sane_readahead(unsigned long nr)
254 return min(nr, (node_page_state(numa_node_id(), NR_INACTIVE_FILE)
255 + node_page_state(numa_node_id(), NR_FREE_PAGES)) / 2);
259 * Submit IO for the read-ahead request in file_ra_state.
261 static unsigned long ra_submit(struct file_ra_state *ra,
262 struct address_space *mapping, struct file *filp)
266 actual = __do_page_cache_readahead(mapping, filp,
267 ra->start, ra->size, ra->async_size);
273 * Set the initial window size, round to next power of 2 and square
274 * for small size, x 4 for medium, and x 2 for large
275 * for 128k (32 page) max ra
276 * 1-8 page = 32k initial, > 8 page = 128k initial
278 static unsigned long get_init_ra_size(unsigned long size, unsigned long max)
280 unsigned long newsize = roundup_pow_of_two(size);
282 if (newsize <= max / 32)
283 newsize = newsize * 4;
284 else if (newsize <= max / 4)
285 newsize = newsize * 2;
293 * Get the previous window size, ramp it up, and
294 * return it as the new window size.
296 static unsigned long get_next_ra_size(struct file_ra_state *ra,
299 unsigned long cur = ra->size;
300 unsigned long newsize;
307 return min(newsize, max);
311 * On-demand readahead design.
313 * The fields in struct file_ra_state represent the most-recently-executed
316 * |<----- async_size ---------|
317 * |------------------- size -------------------->|
318 * |==================#===========================|
319 * ^start ^page marked with PG_readahead
321 * To overlap application thinking time and disk I/O time, we do
322 * `readahead pipelining': Do not wait until the application consumed all
323 * readahead pages and stalled on the missing page at readahead_index;
324 * Instead, submit an asynchronous readahead I/O as soon as there are
325 * only async_size pages left in the readahead window. Normally async_size
326 * will be equal to size, for maximum pipelining.
328 * In interleaved sequential reads, concurrent streams on the same fd can
329 * be invalidating each other's readahead state. So we flag the new readahead
330 * page at (start+size-async_size) with PG_readahead, and use it as readahead
331 * indicator. The flag won't be set on already cached pages, to avoid the
332 * readahead-for-nothing fuss, saving pointless page cache lookups.
334 * prev_pos tracks the last visited byte in the _previous_ read request.
335 * It should be maintained by the caller, and will be used for detecting
336 * small random reads. Note that the readahead algorithm checks loosely
337 * for sequential patterns. Hence interleaved reads might be served as
340 * There is a special-case: if the first page which the application tries to
341 * read happens to be the first page of the file, it is assumed that a linear
342 * read is about to happen and the window is immediately set to the initial size
343 * based on I/O request size and the max_readahead.
345 * The code ramps up the readahead size aggressively at first, but slow down as
346 * it approaches max_readhead.
350 * A minimal readahead algorithm for trivial sequential/random reads.
353 ondemand_readahead(struct address_space *mapping,
354 struct file_ra_state *ra, struct file *filp,
355 bool hit_readahead_marker, pgoff_t offset,
356 unsigned long req_size)
358 int max = ra->ra_pages; /* max readahead pages */
363 * It's the expected callback offset, assume sequential access.
364 * Ramp up sizes, and push forward the readahead window.
366 if (offset && (offset == (ra->start + ra->size - ra->async_size) ||
367 offset == (ra->start + ra->size))) {
368 ra->start += ra->size;
369 ra->size = get_next_ra_size(ra, max);
370 ra->async_size = ra->size;
374 prev_offset = ra->prev_pos >> PAGE_CACHE_SHIFT;
375 sequential = offset - prev_offset <= 1UL || req_size > max;
378 * Standalone, small read.
379 * Read as is, and do not pollute the readahead state.
381 if (!hit_readahead_marker && !sequential) {
382 return __do_page_cache_readahead(mapping, filp,
383 offset, req_size, 0);
387 * Hit a marked page without valid readahead state.
388 * E.g. interleaved reads.
389 * Query the pagecache for async_size, which normally equals to
390 * readahead size. Ramp it up and use it as the new readahead size.
392 if (hit_readahead_marker) {
396 start = radix_tree_next_hole(&mapping->page_tree, offset,max+1);
399 if (!start || start - offset > max)
403 ra->size = start - offset; /* old async_size */
404 ra->size = get_next_ra_size(ra, max);
405 ra->async_size = ra->size;
411 * - first read on start of file
412 * - sequential cache miss
413 * - oversize random read
414 * Start readahead for it.
417 ra->size = get_init_ra_size(req_size, max);
418 ra->async_size = ra->size > req_size ? ra->size - req_size : ra->size;
421 return ra_submit(ra, mapping, filp);
425 * page_cache_sync_readahead - generic file readahead
426 * @mapping: address_space which holds the pagecache and I/O vectors
427 * @ra: file_ra_state which holds the readahead state
428 * @filp: passed on to ->readpage() and ->readpages()
429 * @offset: start offset into @mapping, in pagecache page-sized units
430 * @req_size: hint: total size of the read which the caller is performing in
433 * page_cache_sync_readahead() should be called when a cache miss happened:
434 * it will submit the read. The readahead logic may decide to piggyback more
435 * pages onto the read request if access patterns suggest it will improve
438 void page_cache_sync_readahead(struct address_space *mapping,
439 struct file_ra_state *ra, struct file *filp,
440 pgoff_t offset, unsigned long req_size)
447 ondemand_readahead(mapping, ra, filp, false, offset, req_size);
449 EXPORT_SYMBOL_GPL(page_cache_sync_readahead);
452 * page_cache_async_readahead - file readahead for marked pages
453 * @mapping: address_space which holds the pagecache and I/O vectors
454 * @ra: file_ra_state which holds the readahead state
455 * @filp: passed on to ->readpage() and ->readpages()
456 * @page: the page at @offset which has the PG_readahead flag set
457 * @offset: start offset into @mapping, in pagecache page-sized units
458 * @req_size: hint: total size of the read which the caller is performing in
461 * page_cache_async_ondemand() should be called when a page is used which
462 * has the PG_readahead flag; this is a marker to suggest that the application
463 * has used up enough of the readahead window that we should start pulling in
467 page_cache_async_readahead(struct address_space *mapping,
468 struct file_ra_state *ra, struct file *filp,
469 struct page *page, pgoff_t offset,
470 unsigned long req_size)
477 * Same bit is used for PG_readahead and PG_reclaim.
479 if (PageWriteback(page))
482 ClearPageReadahead(page);
485 * Defer asynchronous read-ahead on IO congestion.
487 if (bdi_read_congested(mapping->backing_dev_info))
491 ondemand_readahead(mapping, ra, filp, true, offset, req_size);
493 EXPORT_SYMBOL_GPL(page_cache_async_readahead);